Risk prediction of PLA2R-Ab-negative membranous nephropathy: an interpretable multicenter machine learning model.

BACKGROUND: Approximately 10%-30% patients with biopsy-proven membranous nephropathy (MN) are seronegative for anti-phospholipase A2 receptor antibody (PLA2R-Ab). Only ~5% are truly non-PLA2R MN and are associated with alternative antigens (e.g., THSD7A), while the majority represent false-negative PLA2R-associated MN due to insufficient assay sensitivity. Accurate noninvasive diagnostic tools for this population are lacking. METHODS: This multicenter retrospective risk stratification study enrolled 692 PLA2R Ab negative patients in the derivation cohort and 333 patients in an independent external validation cohort. We developed and validated an interpretable VotingSoft machine learning model based on 13 key clinical variables. RESULTS: The model achieved excellent risk stratification performance: AUC 0.878 (95% confidence interval (CI): 0.850-0.903) in five-fold cross-validation, 0.894 (95% CI: 0.854-0.930) in the internal test set, and 0.905 (95% CI: 0.867-0.935) in the external validation cohort. The model showed a specificity of 0.819, sensitivity of 0.764, and negative predictive value of 0.876. Key predictors were PLA2R Ab level, eGFR, age, and serum albumin. CONCLUSION: This multicenter interpretable machine learning model provides an auxiliary tool for risk stratification of PLA2R-Ab-negative membranous nephropathy (MN) using routine clinical indicators. Its high specificity may facilitate early risk stratification; however, renal biopsy remains the gold standard for definitive pathological diagnosis.